Uploaded by mail

Satellite Image Classification with Deep Learning Survey

International Journal of Trend in Scientific Research and Development (IJTSRD)
Volume 4 Issue 2, February 2020 Available Online: www.ijtsrd.com e-ISSN: 2456 – 6470
Satellite Image Classification with Deep Learning: Survey
Roshni Rajendran1, Liji Samuel2
Tech Student, 2Assistant Professor,
1,2Department of Computer Science and Engineering, Sree Buddha College of Engineering, Kerala, India
How to cite this paper: Roshni Rajendran
| Liji Samuel "Satellite Image Classification
with Deep Learning: Survey" Published in
International Journal
of Trend in Scientific
Development (ijtsrd),
Volume-4 | Issue-2,
Satellite imagery is important for many applications including disaster
response, law enforcement and environmental monitoring etc. These
applications require the manual identification of objects in the imagery.
Because the geographic area to be covered is very large and the analysts
available to conduct the searches are few, thus an automation is required. Yet
traditional object detection and classification algorithms are too inaccurate
and unreliable to solve the problem. Deep learning is a part of broader family
of machine learning methods that have shown promise for the automation of
such tasks. It has achieved success in image understanding by means that of
convolutional neural networks. The problem of object and facility recognition
in satellite imagery is considered. The system consists of an ensemble of
convolutional neural networks and additional neural networks that integrate
satellite metadata with image features.
Copyright © 2019 by author(s) and
International Journal of Trend in Scientific
Research and Development Journal. This
is an Open Access article distributed
under the terms of
Commons Attribution
KEYWORDS: Deep Learning, Convolutional neural network, VGG, GPU
Deep learning is a class of machine learning models that
represent data at different levels of abstraction by means of
multiple processing layers. It has achieved astonishing
success in object detection and classification by combining
large neural network models, called CNN with powerful GPU.
CNN-based algorithms have dominated the annual ImageNet
Large Scale Visual Recognition Challenge for detecting and
classifying objects in photographs. This success has caused a
revolution in image understanding, and the major
technology companies, including Google, Microsoft and
Facebook, have already deployed CNN-based products and
A CNN consists of a series of processing layers as shown in
Fig. 1.1. Each layer is a family of convolution filters that
detect image features [1]. Near the end of the series, the CNN
combines the detector outputs in fully connected “dense”
layers, finally producing a set of predicted probabilities, one
for each class. The network itself learns which features to
detect, and how to detect them, as it trains.
Fig:1 Architecture of satellite image classification
Unique Paper ID – IJTSRD30031
The earliest successful CNNs consisted of less than 10 layers
and were designed for handwritten zip code recognition.
LeNet had 5 layers and AlexNet had 8 layers. Since then, the
trend has been toward more complexity. VGG appeared with
16 layers followed by Google’s Inception with 22 layers.
Subsequent versions of Inception have even more layers,
ResNet has 152 layers and DenseNet has 161 layers.
Such large CNNs require computational power, which is
provided by advanced GPUs. Open source deep learning
software libraries such as Tensor Flow and Keras, along with
fast GPUs, have helped fuel continuing advances in deep
Various methods are used for identifying the objects in
satellite images. Some of them are discussed below.
1. Rocket Image Classification Based on Deep
Convolutional Neural Network
Liang Zhang et al. [2] says that in the field of aerospace
measurement and control field, optical equipment generates
a large amount of data as image. Thus, it has great research
value for how to process a huge number of image data
quickly and effectively. With the development of deep
learning, great progress has been made in the task of image
classification. The task images are generated by optical
measurement equipment are classified using the deep
learning method. Firstly, based on residual network, a
general deep learning image classification framework, a
binary image classification network namely rocket image
Volume – 4 | Issue – 2
January-February 2020
Page 583
International Journal of Trend in Scientific Research and Development (IJTSRD) @ www.ijtsrd.com eISSN: 2456-6470
and other image is built. Secondly, on the basis of the binary
cross entropy loss function, the modified loss function is
used to achieves a better generalization effect on those
images difficult to classify. Then, the visible image data
downloaded from optical equipment is randomly divided
into training set, validation set and test set. The data
augmentation method is used to train the binary
classification model on a relatively small training set. The
optimal model weight is selected according to the loss value
on the validation set. This method has certain value for
exploring the application of deep learning method in the
intelligent and rapid processing of optical equipment task
image in aerospace measurement and control field.
2. Super pixel Partitioning of Very High Resolution
Satellite Images for Large-Scale Classification
Perspectives with Deep Convolutional Neural
T. Postadjiana et al. [3] proposes that supervised
classification is the basic task for landcover map generation.
From semantic segmentation to speech recognition deep
neural networks has outperformed the state-of-the-art
classifiers in many machine learning challenges. Such
strategies are now commonly employed in the literature for
the purpose of land-cover mapping. The system develops the
strategy for the use of deep networks to label very high
resolution satellite images, with the perspective of mapping
regions at country scale. Therefore, a super pixel based
method is introduced in order to (i) ensure correct
delineation of objects and (ii) perform the classification in a
dense way but with decent computing times.
3. Cloud Cover Assessment in Satellite Images via Deep
Ordinal Classification
Chaomin Shen et al. [4] discuss that the percentage of cloud
cover is one of the key indices for satellite imagery analysis.
To date, cloud cover assessment has performed manually in
most ground stations. To facilitate the process, a deep
learning approach for cloud cover assessment in quicklook
satellite images is proposed. Same as the manual operation,
given a quicklook image, the algorithm returns 8 labels
ranging from A to E and *, indicating the cloud percentages
in different areas of the image. This is achieved by
constructing 8 improved VGG-16 models, where parameters
such as the loss function, learning rate and dropout are
tailored for better performance. The procedure of manual
assessment can be summarized as follows. First, determine
whether there is cloud cover in the scene by visual
inspection. Some prior knowledge, e.g., shape, color and
shadow, may be used. Second, estimate the percentage of
cloud presence. Although in reality, the labels are often
determined as follows. If there is no cloud, then A; If a very
small amount of clouds exist, then B; C and D are given to
escalating levels of clouds; and E is given when the whole
part is almost covered by clouds. There is also a label * for
no-data. This mostly happens when the sensor switches,
causing no data for several seconds. The disadvantages of
manual assessment are obvious. First of all, it is tedious
work. Second, results may be inaccurate due to subjective
A novel deep learning application in remote sensing by
assessing the cloud cover in an optical scene is considered.
This algorithm uses parallel VGG 16 networks and returns 8
labels indicating the percentage of cloud cover in respective
Unique Paper ID – IJTSRD30031
subscenes. The proposed algorithm combines several stateof-the-art techniques and achieves reasonable results.
4. Learning Multiscale Deep Features for HighResolution Satellite Image Scene Classification
Qingshan Liu et al. [5] discuss about a multiscale deep
feature learning method for high-resolution satellite image
scene classification. However, satellite images with high
spatial resolution pose many challenging issues in image
classification. First, the enhanced resolution brings more
details; thus, simple lowlevel features (e.g., intensity and
textures) widely used in the case of low-resolution images
are insufficient in capturing efficiently discriminative
information. Second, objects in the same type of scene might
have different scales and orientations. Besides, highresolution satellite images often consist of many different
semantic classes, which makes further classification more
difficult. Taking the commercial scene comprises roads,
buildings, trees, parking lots, and so on. Thus, developing
effective feature representations is critical for solving these
Specifically, at first warp the original satellite image into
multiple different scales. The images in each scale are
employed to train a deep convolutional neural network
(DCNN). However, simultaneously training multiple DCNNs
is time consuming. To address this issue, a DCNN with spatial
pyramid pooling is explored. Since different SPP-nets have
the same number of parameters, which share the identical
initial values, and only fine-tuning the parameters in fully
connected layers ensures the effectiveness of each network,
thereby greatly accelerating the training process. Then, the
multiscale satellite images are fed into their corresponding
SPP-nets, respectively, to extract multiscale deep features.
Finally, a multiple kernel learning method is developed to
automatically learn the optimal combination of such
5. Domain Adaptation for Large Scale Classification of
Very High Resolution Satellite Images With Deep
Convolutional Neural Networks
T. Postadjiana et al. [6] discuss about semantic segmentation
of remote sensing images enables in particular land-cover
map generation for a given set of classes. Very recent
literature has shown the superior performance of DCNN for
many tasks, from object recognition to semantic labelling,
including the classification of VHR satellite images. A simple
yet effective architecture is DCNN. Input images are of size of
65*65*4 (number of bands). Only convolutions of size 3*3
are used to limit the number of parameters. The second line
of the table displays the number of filters per convolution
layer. The ReLU activation function is set after each
convolution in order to introduce non-linearity and maxpooling layers increase the receptive field (the spatial
information taken into account by the filter). Finally, a fullyconnected layer on top sums up the information contained in
all features in the last convolution layer.
However, while plethora of works aim at improving object
delineation on geographically restricted areas, few tend to
solve this classification task at very large scales. New issues
occur such as intra-class class variability, diachrony between
surveys, and the appearance of new classes in a specific area,
that do not exist in the predefined set of labels. Therefore,
this work intends to (i) perform large scale classification and
Volume – 4 | Issue – 2
January-February 2020
Page 584
International Journal of Trend in Scientific Research and Development (IJTSRD) @ www.ijtsrd.com eISSN: 2456-6470
to (ii) expand a set of land-cover classes, using the off-theshelf model learnt in a specific area of interest and adapting
it to unseen areas.
6. Introducing Eurosat: A Novel Dataset and Deep
Learning Benchmark for Land Use and Land Cover
Patrick Helber et al. [7] discuss about the challenge of land
use and land cover classification using Sentinel-2 satellite
images. The key contributions are as follows. A novel dataset
based on Sentinel-2 satellite images covering 13 different
spectral bands and consisting of 10 classes with in total
27,000 labeled images are presented. The state-of-the-art
CNN on this novel dataset with its different spectral bands
are considered. Also evaluate deep CNNs on existing remote
sensing datasets. With the proposed novel dataset, a better
overall classification accuracy is gained. The classification
system resulting from the proposed research opens a gate
towards various Earth observation applications.
The challenge of land use and land cover classification is
considered. For this task, dataset based on Sentinel-2
satellite images are used. The proposed dataset consists of
10 classes covering 13 different spectral bands with in total
27,000 labeled images. The evaluated state of the art deep
CNNs on this novel dataset. Also evaluated deep CNNs on
existing remote sensing datasets and compared the obtained
results. For the novel dataset, the performance is analysed
based on different spectral bands. The proposed research
can be leveraged for multiple real-world Earth observation
applications. Possible applications are land use and land
cover change detection and the improvment of geographical
7. Cloud Classification of Satellite Image Based on
Convolutional Neural Networks
Keyang Cai et al. [8] proposes about cloud classification of
satellite image in meteorological forecast. Traditional
machine learning methods need to manually design and
extract a large number of image features, while the
utilization of satellite image features is not high. CNN is the
first truly successful learning algorithm for multi-layer
network structure. Compared with the general forward BP
algorithm, CNN can reduce the number of parameters
needed in learning in spatial relations, so as to improve the
training performance. A small piece of local feel in the
picture as the bottom of the hierarchical structure of the
input, continue to transfer to the next layer, each layer
through the digital filter to obtain the characteristics of the
data. This method has significant effect on the observed data
such as scaling, rotation and so on. The cloud classification
process of satellite based on deep convolution neural
network includes pre-processing, feature extraction,
classification and other steps. A convolution neural network
for cloud classification, which can automatically learn
features and obtain classification results is constructed. The
method has high precision and good robustness.
8. Hyperspectral Classification Using Stacked
Autoencoders With Deep Learning
A. Okan Bilge Ozdemir et al. [9] says that the stacked
autoencoders which are widely utilized in deep learning
research are applied to remote sensing domain for
hyperspectral classification. High dimensional hyperspectral
data is an excellent candidate for deep learning methods.
Unique Paper ID – IJTSRD30031
However, there are no works in literature that focuses on
such deep learning approaches for hyperspectral imagery.
This aims to fill this gap by utilizing stacked autoencoders.
Using stacked autoencoders, intrinsic representations of the
data are learned in an unsupervised way. Using labeled data,
these representations are fine tuned. Then, using a soft-max
activation function, hyperspectral classification is done.
Parameter optimization of Stacked Autoencoders (SAE) is
done with extensive experiments. It focuses on utilization of
representation learning methods for hyperspectral
classification task on high definition hyperspectral images.
Strong side of representation learning methods are its
unsupervised automatic feature learning step which makes it
possible to omit the feature extraction step that requires
domain knowledge.
Proposed method utilizes SAEs to learn discriminative
representations in the data. Using the label information in
the training set, representations are fine tuned. Then, the
test set is classified using these representations. Effects of
different parameters on the performance results are also
investigated. Most effective parameter is found as the epoch
number of the supervised classification step. For each
parameter (learning rate, batch size and epoch numbers),
optimal values are approximated with extensive testing.
From 432 different parameter configurations, optimal
parameter values are found as 25, 60, 100 and 5 for batch
size, epoch numbers (representation & classification) and
learning rate, respectively. This score is greatly promising
when compared to state-of-the-art methods. This study can
be enhanced by two important additions, neighbourhood
factors and utilization of unlabelled data. With these
additions as future study, more general and robust results
can be achieved.
9. Deep Learning Crop Classification Approach Based
on Sparse Coding of Time Series of Satellite Data
Mykola Lavreniuk et al. [10] says that crop classification
maps based on high resolution remote sensing data are
essential for supporting sustainable land management. The
most challenging problems for their producing are collecting
of ground-based training and validation datasets, nonregular satellite data acquisition and cloudiness. To increase
the efficiency of ground data utilization it is important to
develop classifiers able to be trained on the data collected in
the previous year. A deep learning method is analyzed for
providing crop classification maps using in-situ data that has
been collected in the previous year. Main idea is to utilize
deep learning approach based on sparse autoencoder. At the
first stage it is trained on satellite data only and then neural
network fine-tuning is conducted based on in-situ data form
the previous year. Taking into account that collecting ground
truth data is very time consuming and challenging task, the
proposed approach allows us to avoid necessity for annual
collecting in-situ data for the same territory.
This approach is based on converting time series of
observations into sparse vectors from the unified
hyperspace with similar representation. This idea allows
further utilization of the autoencoder to learn dependencies
in data without labels and fine-tuning final neural network
using available in-situ data from previous year. This
methodology was utilized to provide crop classification map
for central part of Ukraine for 2017 based on multi-temporal
Sentinel-1 images with different acquisition dates.
Volume – 4 | Issue – 2
January-February 2020
Page 585
International Journal of Trend in Scientific Research and Development (IJTSRD) @ www.ijtsrd.com eISSN: 2456-6470
10. Deep Learning for Amazon Satellite Image Analysis
Lior Bragilevsky et al. [11] says that satellite image analysis
has become very important in a lot of areas such as
monitoring floods and other natural disasters, earthquake
and tsunami prediction, ship tracking and navigation,
monitoring the effects of climate change etc. Large size and
inaccessibility of some of the regions of Amazon wants best
satellite imaging system. Useful informations are extracted
from satellite images will help to observe and understand
the changing nature of the Amazon basin, and help better to
manage deforestation and its consequences. Some images
were labelled and could be used to train various machine
learning algorithms. Another set of test image was provided
without labels and was used to make predictions. The
predicted labels for test images were submitted to the
competition for scoring. The hope was that well-trained
models arising from this competition would allow for a
better understanding of the deforestation process in the
Amazon basin and give insight on how to better manage this
Machine learning can be the key to saving the world from
losing football field-sized forest areas each second. As
deforestation in the Amazon basin causes devastating effects
both on the ecosystem and the environment, there is urgent
need to better understand and manage its changing
landscape. A competition was recently conducted to develop
algorithms to analyze satellite images of the Amazon.
Successful algorithms will be able to detect subtle features in
different image scenes, giving us the crucial data needed to
be able to manage deforestation and its consequences more
11. Deep Learning-Based Large-Scale Automatic
Satellite Crosswalk Classification
Rodrigo F. Berriel et al. [12] proposes about a zebra crossing
classification and detection system. It is one of the main
tasks for mobility autonomy. Although important, there are
few data available on where crosswalks are in the world. The
automatic annotation of crosswalks locations worldwide can
be very useful for online maps, GPS applications and many
others. In addition, the availability of zebra crossing
locations on these applications can be of great use to people
with disabilities, to road management, and to autonomous
vehicles. However, automatically annotating this kind of data
is a challenging task. They are often aging (painting fading
away), occluded by vehicle and pedestrians, darkened by
strong shadows, and many other factors.
High-resolution satellite imagery has been increasingly used
on remote sensing classification problems. One of the main
factors is the availability of this kind of data. Despite the high
availability, very little effort has been placed on the zebra
crossing classification problem. Crowd sourcing systems are
exploited in order to enable the automatic acquisition and
annotation of a large-scale satellite imagery database for
crosswalks related tasks.
Various methods to capture and identify the objects in the
satellite imageries are analyzed. Each of them works with
different schemes. Each scheme varies on the convolution
layers used and also their application used.
Unique Paper ID – IJTSRD30031
A deep learning system that classifies objects in the satellite
imagery is considered. The system consists of an ensemble of
CNNs with post-processing neural networks that combine
the predictions from the CNNs with satellite metadata.
Combined with a detection component, this system could
search large amounts of satellite imagery for objects or
facilities of interest. Analyzing satellite imagery has very
important role in various applications such as it could help
law enforcement officers to detect unlicensed mining
operations or illegal fishing vessels, assist natural disaster
response teams with the mapping of mud slides or hurricane
damage and enable investors to monitor crop growth or oil
well development more effectively.
[1] M. Pritt and G. Chern, "Satellite Image Classification
with Deep Learning," 2017 IEEE Applied Imagery
Pattern Recognition Workshop (AIPR), Washington, DC,
2017, pp. 1-7.
[2] L. Zhang, Z. Chen, J. Wang and Z. Huang, "Rocket Image
Classification Based on Deep Convolutional Neural
Network," 2018 10th International Conference on
Communications, Circuits and Systems (ICCCAS),
Chengdu, China, 2018, pp. 383-386.
[3] C. Shen, C. Zhao, M. Yu and Y. Peng, "Cloud Cover
Assessment in Satellite Images Via Deep Ordinal
Classification," IGARSS 2018 - 2018 IEEE International
Geoscience and Remote Sensing Symposium, Valencia,
2018, pp. 3509-3512.
[4] T. Postadjian, A. L. Bris, C. Mallet and H. Sahbi,
"Superpixel Partitioning of Very High Resolution
Satellite Images for Large-Scale Classification
Perspectives with Deep Convolutional Neural
Networks," IGARSS 2018 - 2018 IEEE International
Geoscience and Remote Sensing Symposium, Valencia,
2018, pp. 1328-1331.
[5] Q. Liu, R. Hang, H. Song and Z. Li, "Learning Multiscale
Deep Features for High-Resolution Satellite Image
Scene Classification," in IEEE Transactions on
Geoscience and Remote Sensing, vol. 56, no. 1, pp. 117126, Jan. 2018.
[6] T. Postadjian, A. L. Bris, H. Sahbi and C. Malle, "Domain
Adaptation for Large Scale Classification of Very High
Resolution Satellite Images with Deep Convolutional
Neural Networks," IGARSS 2018 - 2018 IEEE
International Geoscience and Remote Sensing
Symposium, Valencia, 2018, pp. 3623-3626.
[7] P. Helber, B. Bischke, A. Dengel and D. Borth,
"Introducing Eurosat: A Novel Dataset and Deep
Learning Benchmark for Land Use and Land Cover
Classification," IGARSS 2018 - 2018 IEEE International
Geoscience and Remote Sensing Symposium, Valencia,
2018, pp. 204-207.
[8] K. Cai and H. Wang, "Cloud classification of satellite
image based on convolutional neural networks," 2017
8th IEEE International Conference on Software
Engineering and Service Science (ICSESS), Beijing, 2017,
pp. 874-877.
[9] A. O. B. Özdemir, B. E. Gedik and C. Y. Y. Çetin,
autoencoders with deep learning," 2014 6th Workshop
Volume – 4 | Issue – 2
January-February 2020
Page 586
International Journal of Trend in Scientific Research and Development (IJTSRD) @ www.ijtsrd.com eISSN: 2456-6470
on Hyperspectral Image and Signal Processing:
Evolution in Remote Sensing (WHISPERS), Lausanne,
2014, pp. 1-4.
[10] M. Lavreniuk, N. Kussul and A. Novikov, "Deep Learning
Crop Classification Approach Based on Sparse Coding
of Time Series of Satellite Data," IGARSS 2018 - 2018
IEEE International Geoscience and Remote Sensing
Symposium, Valencia, 2018, pp. 4812-4815.
Unique Paper ID – IJTSRD30031
[11] L. Bragilevsky and I. V. Bajić, "Deep learning for
Amazon satellite image analysis," 2017 IEEE Pacific Rim
Conference on Communications, Computers and Signal
Processing (PACRIM), Victoria, BC, 2017, pp. 1-5.
[12] R. F. Berriel, A. T. Lopes, A. F. de Souza and T. OliveiraSantos, "Deep Learning-Based Large-Scale Automatic
Satellite Crosswalk Classification," in IEEE Geoscience
and Remote Sensing Letters, vol. 14, no. 9, pp. 15131517, Sept. 2017.
Volume – 4 | Issue – 2
January-February 2020
Page 587